Analyzing the Performance of Univariate and Multivariate Machine Learning Models in Soil Movement Prediction: A Comparative Study

Movement of soil and associated landslides frequently occur in hilly areas. Regular monitoring, accurate prediction, and timely alerting of people about soil movements on hills susceptible to landslides are essential due to the potential destruction to life and property. A more recent strategy for predicting soil movement is the use of machine learning (ML) models. Different univariate and multivariate ML models have been proposed in the literature. However, evaluating these univariate and multivariate approaches in predicting real-world landslides have received less attention. This paper’s primary goal is to develop and compare the ability of univariate and multivariate ML models (Autoregression (AR), Seasonal Autoregressive Integrated Moving Average (SARIMA), Sequential Minimal Optimization regression (SMOreg), Multilayer Perceptron (MLP), and Long-short Term Memory (LSTM)) to predict movements at a real-world landslide. The case study used for the analysis in this paper is the Tangni landslide in India. This study makes use of weekly averaged soil movement data collected from June 2012 to December 2013 (78 weeks) at the Tangni landslide site. The dataset comprises measurements from five sensors. To calibrate the parameters in each model, we divided the collected data into a training dataset (first 62 weeks) and a test dataset (last 16 weeks). Performance analysis of the models utilized Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-squared values. The training results revealed that the univariate AR model demonstrated the best performance, achieving an RMSE of 0.149 degrees and an R-squared value of 0.572. The univariate SMOreg model obtained the second-best performance with an RMSE of 0.336 degrees and an R-squared value of 0.582. However, on the test dataset, the multivariate SARIMAX model outperformed the other models, achieving an RMSE of 0.351 degrees and an R-squared value of 0.769. The univariate SARIMA model also performed well with an RMSE of 0.356 degrees and an R-squared value of 0.741. The findings of this study can have significant implications in the field of landslide prediction and prevention. The results indicate that the multivariate SARIMAX model, the most accurate in predicting soil movements, can aid in developing early warning systems against landslides in hilly areas.